NonuniformFFTs.jl

Yet another package for computing multidimensional non-uniform fast Fourier transforms (NUFFTs) in Julia.

Like other existing packages, computations are parallelised using threads. By default, all available Julia threads are used.

Basic usage

Type-1 (or adjoint) NUFFT in one dimension

using NonuniformFFTs

N = 256   # number of Fourier modes
Np = 100  # number of non-uniform points

# Generate some non-uniform random data
T = Float64                # non-uniform data is real (can also be complex)
xp = rand(T, Np) .* T(2π)  # non-uniform points in [0, 2π]
vp = randn(T, Np)          # random values at points

# Create plan for data of type T
plan_nufft = PlanNUFFT(T, N; m = HalfSupport(8))  # larger support increases accuracy

# Set non-uniform points
set_points!(plan_nufft, xp)

# Perform type-1 NUFFT on preallocated output
ûs = Array{Complex{T}}(undef, size(plan_nufft))
exec_type1!(ûs, plan_nufft, vp)

Type-2 (or direct) NUFFT in one dimension

using NonuniformFFTs

N = 256   # number of Fourier modes
Np = 100  # number of non-uniform points

# Generate some uniform random data
T = Float64                        # non-uniform data is real (can also be complex)
xp = rand(T, Np) .* T(2π)          # non-uniform points in [0, 2π]
ûs = randn(Complex{T}, N ÷ 2 + 1)  # random values at points (we need to store roughly half the Fourier modes for complex-to-real transform)

# Create plan for data of type T
plan_nufft = PlanNUFFT(T, N; m = HalfSupport(8))

# Set non-uniform points
set_points!(plan_nufft, xp)

# Perform type-2 NUFFT on preallocated output
vp = Array{T}(undef, Np)
exec_type2!(vp, plan_nufft, ûs)

More examples

Multidimensional transforms

using NonuniformFFTs
using StaticArrays: SVector  # for convenience

Ns = (256, 256)  # number of Fourier modes in each direction
Np = 1000        # number of non-uniform points

# Generate some non-uniform random data
T = Float64                                      # non-uniform data is real (can also be complex)
d = length(Ns)                                   # number of dimensions (d = 2 here)
xp = [T(2π) * rand(SVector{d, T}) for _ ∈ 1:Np]  # non-uniform points in [0, 2π]ᵈ
vp = randn(T, Np)                                # random values at points

# Create plan for data of type T
plan_nufft = PlanNUFFT(T, Ns; m = HalfSupport(8))

# Set non-uniform points
set_points!(plan_nufft, xp)

# Perform type-1 NUFFT on preallocated output
ûs = Array{Complex{T}}(undef, size(plan_nufft))
exec_type1!(ûs, plan_nufft, vp)

# Perform type-2 NUFFT on preallocated output
exec_type2!(vp, plan_nufft, ûs)

Multiple transforms on the same non-uniform points

using NonuniformFFTs

N = 256   # number of Fourier modes
Np = 100  # number of non-uniform points
ntrans = Val(3)  # number of simultaneous transforms

# Generate some non-uniform random data
T = Float64                # non-uniform data is real (can also be complex)
xp = rand(T, Np) .* T(2π)  # non-uniform points in [0, 2π]
vp = ntuple(_ -> randn(T, Np), ntrans)  # random values at points (one vector per transformed quantity)

# Create plan for data of type T
plan_nufft = PlanNUFFT(T, N; ntransforms = ntrans)

# Set non-uniform points
set_points!(plan_nufft, xp)

# Perform type-1 NUFFT on preallocated output (one array per transformed quantity)
ûs = ntuple(_ -> Array{Complex{T}}(undef, size(plan_nufft)), ntrans)
exec_type1!(ûs, plan_nufft, vp)

# Perform type-2 NUFFT on preallocated output (one vector per transformed quantity)
vp_interp = map(similar, vp)
exec_type2!(vp, plan_nufft, ûs)

Using the AbstractNFFTs.jl interface

This package also implements the AbstractNFFTs.jl interface as an alternative API for constructing plans and evaluating transforms. This can be useful for comparing with similar packages such as NFFT.jl.

using NonuniformFFTs
using AbstractNFFTs
using LinearAlgebra: mul!

Ns = (256, 256)  # number of Fourier modes in each direction
Np = 1000        # number of non-uniform points

# Generate some non-uniform random data
T = Float64                      # must be a real data type (Float32, Float64)
d = length(Ns)                   # number of dimensions (d = 2 here)
xp = rand(T, (d, Np)) .- T(0.5)  # non-uniform points in [-1/2, 1/2)ᵈ; must be given as a (d, Np) matrix
vp = randn(Complex{T}, Np)       # random values at points (must be complex)

# Create plan for data of type Complex{T}. Note that we pass the points `xp` as
# a first argument, which calls an AbstractNFFTs-compatible constructor.
p = PlanNUFFT(xp, Ns)

# Getting the expected dimensions of input and output data.
AbstractNFFTs.size_in(p)   # (256, 256)
AbstractNFFTs.size_out(p)  # (1000,)

# Perform adjoint NFFT, a.k.a. type-1 NUFFT (non-uniform to uniform)
us = adjoint(p) * vp      # allocates output array `us`
mul!(us, adjoint(p), vp)  # uses preallocated output array `us`

# Perform forward NFFT, a.k.a. type-2 NUFFT (uniform to non-uniform)
wp = p * us
mul!(wp, p, us)

# Setting a different set of non-uniform points
AbstractNFFTs.nodes!(p, xp)

Note: the AbstractNFFTs.jl interface currently only supports complex-valued non-uniform data. For real-to-complex transforms, the NonuniformFFTs.jl API demonstrated above should be used instead.

Differences with other packages

This package roughly follows the same notation and conventions of the FINUFFT library and its Julia interface, with a few differences detailed below.

Conventions used by this package

We try to preserve as much as possible the conventions used in FFTW3. In particular, this means that:

The FFT outputs are ordered starting from mode $k = 0$ to $k = N/2 - 1$ (for even $N$) and then from $-N/2$ to $-1$. Wavenumbers can be obtained in this order by calling AbstractFFTs.fftfreq(N, N). Use AbstractFFTs.fftshift to get Fourier modes in increasing order $-N/2, …, -1, 0, 1, …, N/2 - 1$. In FINUFFT, one should set modeord = 1 to get this order.
The type-1 NUFFT (non-uniform to uniform) is defined with a minus sign in the exponential. This is the same convention as the forward DFT in FFTW3. In particular, this means that performing a type-1 NUFFT on uniform points gives the same output than performing a FFT using FFTW3. In FINUFFT, this corresponds to setting iflag = -1 in type-1 transforms. Conversely, type-2 NUFFTs (uniform to non-uniform) are defined with a plus sign, equivalently to the backward DFT in FFTW3.

For compatibility with other packages such as NFFT.jl, these conventions are not applied when the AbstractNFFTs.jl interface is used (see example above). In this specific case, modes are assumed to be ordered in increasing order, and the opposite sign convention is used for Fourier transforms.

Differences with NFFT.jl

This package allows NUFFTs of purely real non-uniform data.
Different convention is used: non-uniform points are expected to be in $[0, 2π]$.

Differences with FINUFFT / FINUFFT.jl

This package is written in "pure" Julia (besides the FFTs themselves which rely on the FFTW3 library, via their Julia interface).
This package allows NUFFTs of purely real non-uniform data. Moreover, transforms can be performed in for an arbitrary number of dimensions.
A different smoothing kernel function is used (backwards Kaiser–Bessel kernel by default).
It is possible to use the same plan for type-1 and type-2 transforms, reducing memory requirements in cases where one wants to perform both.